Migration, Urbanicity, and Health: Mutually reinforcing contextual and compositional place-based determinants of health
Department of Population Health Sciences
2024-11-19
Aims:
- Quantify the associations between place and self-reported mental health at the county level
- Explore how county-level migration can enhance our capacity to understand and explain county-level health
- Identify how a county’s position within a migration system may be related to county-level health
Mutually reinforcing processes that define the health of a place
Mutually reinforcing processes that define the health of a place
Mutually reinforcing processes that define the health of a place
Mutually reinforcing processes that define the health of a place
Mutually reinforcing processes that define the health of a place
Mutually reinforcing processes that define the health of a place
Motivation:
Ha Makoae, Lesotho
Motivation:
Structural Determinants of Health
Motivation:
Urban-rural divide
Motivation:
Looking forward
Inequality
Rural-urban divides account for approximately 40% of within country inequality (Young 2013)
Growth of urban populations is associated with greater rural-urban gaps in health outcomes (Beatriz et al. 2018)
Climate change
Climate-induced migration could result in up to 300% increase in population density (Hsiang and Sobel 2016)
Urbanicity-related inequalities are related to climate susceptibility (Fong, Mehta, and Bell 2020)
Goals:
Place-based determinants of health are often static and do not account for migration / movement of people
Assessments of the impact of migration on health often examine individual-level effects
Aim 1: Hypotheses
Quantify the associations between place and self-reported mental health at the county level
Hypothesis 1A: County-level averages in poor mental health days are related to urbanicity after accounting for county-level demographic differences
Hypothesis 1B: This relationship can be explained by differences in factors linked to the built environment (e.g. access to exercise, mental health providers, air pollution, violent crime, and severe housing problems)
Note:
Aim 1 findings have been published in the Community Mental Health Journal.
Aim 1: Data
All data come from the County Health Rankings and Roadmaps 2021 dataset
Outcome: self-reported poor mental health days from the Behavioral Risk Factor Surveillance System (BRFSS)
Exposure: urbanicity, as defined by the National Center for Health Statistics (NCHS)
Controls: sociodemographic makeup of each county
Aim 1: Data
All data come from the County Health Rankings and Roadmaps 2021 dataset
Aim 1: Approach
Aim 1: Hypothesis 1A Approach
County-level averages in poor mental health days are related to urbanicity after accounting for county-level demographic differences
Calculate propensity scores
- We used logistic regression to calculate the propensity of belonging to a given urbanicity category conditional on confounding variables (education, income, percent Black, percent Hispanic, age) relative to reference category (small metro)
Calculate relative change in poor mental health days
- We used a mixed effects linear regression model with random effect for state (with inverse probability weights)
Aim 1: Hypothesis 1B Approach
The relationship between urbanicity and mental health can be explained by differences in factors linked to the built environment
We investigated 8 potential mediating factors:
Aim 1: Hypothesis 1B Approach
The relationship between urbanicity and mental health can be explained by differences in factors linked to the built environment
Mediation analysis in two steps (Imai, Keele, and Tingley 2010; VanderWeele and Vansteelandt 2014)
Mentally unhealthy days ~ mediator + urbanicity
Mediator ~ urbanicity
Estimate of mediation: effect of urbanicity on mediator * effect of mediator on mentally unhealthy days
Aim 1: Results
Controlling for state, age, income, education, and race/ethnicity, large central metro counties reported 0.24 fewer average poor mental health days than small metro counties (t = -5.78, df = 423, p < 0.001)
Noncore counties had 0.07 more average poor mental health days than small metro counties (t = 3.06, df = 1690, p = 0.002)
Better mental health in large central metro counties was partially mediated by differences in the built environment, such as better food environments. Poorer mental health in noncore counties was not mediated by considered mediators.
Aim 1: Pitfalls and Alternatives
BRFSS data is modeled at the state level
We are using race and ethnicity as proxies for lived experiences that may differ by identity due to systemic injustices
Model dependent
Our findings are statistically significant but not clinically or biologically significant
County-level analyses may not adequately capture neighborhood nuances
Ecological analyses - which matters more: geography or population?
Aim 2: Hypotheses
Explore how county-level migration can enhance our capacity to understand and explain county-level health
Hypothesis 2A: County-to-county migration patterns improve the explainability of autoregressive models of county-level health outcomes
Hypothesis 2B: The role that county-to-county migration flows plays in county-level health outcomes differs signficantly between rural and urban counties
Hypothesis 2C: Taking into account unmeasured factors in county-to-county migration flows improves our ability to explain county-level health outcomes as well as the differential role that migration plays in urban versus rural counties
Aim 2: Variables
Outcome: \(y_{it}\)
County-level age-adjusted premature mortality rate per 100,000 population of county \(i\) at time \(t\)
Premature mortality: any death occurring before age 75
Each \(t\) is a single year from 2012-2019
Source: CDC WONDER Underlying Cause of Death
Aim 2: Variables
Baseline explanatory factor: \(y_{i,t-1}\)
Lagged county-level age-adjusted premature mortality rate of county \(i\) at time \(t-1\) (i.e. the prior year)
Aim 2: Variables
Primary explanatory factor: \(mig_{it}\)
Weighted average accounting for compositional change in a destination county \(i\) at time \(t\)
\[ mig_{it} = \frac{ \sum_{j\ne i} out_{jit} y_{j,t-1} + y_{i,t-1} (pop_{i, t-1} - \sum_{j\ne i} out_{ijt})}{ \sum_{j\ne i} out_{jit} + (pop_{i,t-1} - \sum_{j\ne i} out_{ijt})} \]
\(out_{jit}\) represents the number of movers from county \(j\) to a destination county \(i\) between year \(t-1\) and year \(t\)
- Source: IRS migration data: number of personal exemptions claimed
\(y_{j, t-1}\) is the lagged premature age-adjusted mortality rate of origin county \(j\)
\(pop_{i,t-1}\) is the population under age 75 of county \(i\) in the prior year \(t-1\)
- Source: CDC WONDER Underlying Cause of Death
Aim 2: Variables
Secondary explanatory factor: \(U_i\)
Urbanicity as defined by the NCHS (same as Aim 1), grouped into urban and rural for each county \(i\)
1948 rural counties
1159 urban counties
Aim 2: Variables
Exploratory explanatory factor: \(amig_{it}\)
Adjusted weighted average migration term for county \(i\) in year \(t\)
\[ amig_{it} = \frac{ \sum_{j\ne i} out_{jit} ( y_{j,t-1} + d_{ij}) + y_{i,t-1} (pop_{i, t-1} - \sum_{j\ne i} out_{ijt})}{ \sum_{j\ne i} out_{jit} + (pop_{i,t-1} - \sum_{j\ne i} out_{ijt})}\]
\(d_{ij}\) accounts for health-related selection of movers to county \(i\) from county \(j\)
When \(d_{ij} > 0\) county \(i\) tends to attract less healthy movers
When \(d_{ij} < 0\) county \(i\) tends to attract healthier movers
Aim 2: Models
Baseline Model:
\[ y_{it} = \beta_0 + \beta_{1t} + \beta_{2}y_{t-1,i} + \mu_{i} + \epsilon_{it} \]
where \(\beta_0\) is the intercept
\(\beta_{1t}\) is a coefficient for the effect of each year \(t\)
\(\beta_2\) is a coefficient to capture the effect of lagged premature age-adjusted mortality
\(\mu_i\) is a random intercept for each county \(i\)
and \(\epsilon_{it}\) represents spatial error as defined on the next slide….
Aim 2: Models
Baseline Model: Spatial error
\[ \epsilon_{it} = \lambda W \epsilon_{it} + u_{it} \]
\(\lambda\) is a scalar to represent the magnitude of spatial error
\(W\) is a spatial weights matrix created using the “queen” criterion which considers counties that share any point as neighbors
\(\mu_{it}\) is a random error term for each county and each year
Aim 2: Models
Iterative process:
- Test splines for \(y_{i,t-1}\)
- Test \(mig_{it}\) with and without splines (Hypothesis 2A)
- Test interaction between \(mig_{it}\) and urbanicity (Hypothesis 2B)
- Replace \(mig_{it}\) with \(amig_{it}\) and test values of \(d_{ij}\) incrementing by 50 from -200 to 200 in units of premature deaths per 100,000 population (Hypothesis 2C)
Aim 2: Models
Iterative process:
Test values of \(d_{ij}\) dependent upon the urbanicity category of origin county \(j\) and destination county \(i\) such that \(d_{ij}\) could be one of four values:
- \(d_{uu}\) urban to urban
- \(d_{rr}\) rural to rural
- \(d_{ru}\) rural to urban
- \(d_{ur}\) urban to rural
- Where each of \((d_{uu}, d_{rr}, d_{ru}, d_{ur})\) were varied factorially
- Coarse search: incrementing by 50 from -200 to 200
- Then incrementing by 20 from -100 to 100 around the minimum found in the coarse search
- We tested a total of 13,122 \((= 9^4 * 2)\) combinations of \((d_{uu}, d_{rr}, d_{ru}, d_{ur})\)
Aim 2: Results
Hypothesis 2A: Does the migration term add explainability to the baseline model?
YES - models including the \(mig_{it}\) term had lower BIC scores than models without.
Aim 2: Results
Hypothesis 2B: Does the role of \(mig_{it}\) differ between rural and urban counties?
YES - \(mig_{it}\) significantly enhances model explainability when accounting for urbanicity
Aim 2: Results
Aim 2: Results
Hypothesis 2C: Can we account for unmeasured factors related to health and migration?
Health-related selection may not be important to modeling county-level health. BIC score is minimized when \(d_{ij} = 0\).
Aim 2: Results
Hypothesis 2C: Can we account for unmeasured factors related to health and migration when accounting for urbanicity?
Health-related selection may be important to modeling county-level health when we account for urbanicity. BIC score is minimized when \((d_{uu} = -100, d_{rr} = 0, d_{ru} = 0, d_{ur} = 20)\).
Aim 2: Conclusion
Migration does matter!
Urbanicity matters even after accounting for migration
Some evidence that healthier urban destinations are connected to less healthy urban origins
Aim 2: Pitfalls and Alternatives
Regression to the mean
Premature age-adjusted mortality rates serve as a proxy for overall county-level health
Multicollinearity
IRS data
Internal migration only
Pre-covid
Aim 3: Hypothesis
Is a county’s position within a migration system predictive of county-level health?
Aim 3: Approach
Tensor decomposition
Aim 3: Approach
Constructing the tensor \(X\)
A CUBE!
Origin by destination by time: \(3109\) x \(3109\) x \(8\)
Each entry in the tensor represents the \(log(1+n)\) transformed number of movers from origin \(i\) to destination \(j\) in time period \(k\).
Aim 3: Approach
Tensor Decomposition
Non-Negative Tensor Factorization (NTF) represented as:
\[ X \approx \sum_{r=1}^R o_{ir} \cdot d_{jr} \cdot t_{kr}, \quad \text{such that } o_{ir} \geq 0, \, d_{jr} \geq 0, \, t_{kr} \geq 0 \, \forall \, i, j, k, r. \]
Where each \(rth\) component corresponds to a unique “migration system”, and the loadings \((o_{ir}, d_{jr}, t_{kr})\) represent the significance of each origin county \(i\), destination county \(j\), and year \(k\) to each migration system \(r\).
Aim 3: Component Selection
Aim 3: Three migration systems
Aim 3: Three migration systems
Component 1: Urban to Rural
Aim 3: Three migration systems
Component 2: Specific Phenomena
Aim 3: Three migration systems
Component 3: State borders
Aim 3: Three migration systems
Temporal trends
Aim 3: Connection to county-level health
\[ y_{it} = \beta_0 + \sum_{r=1}^{3} \beta_{o_r} o_{ir} + \sum_{r=1}^{3} \beta_{d_r} d_{jr} + \sum_{r=1}^{3} \beta_{t_r} t_{kr} + \sum_{n=2}^{6} \gamma_n \text{Urbanicity}_n + \mu_{it} \]
Predictor variables: \(o_{ir}, d_{jr}, t_{kr}\)
- Loadings from the tensor decomposition corresponding to origin, destination, and time
Outcome variable: \(y_{it}\)
- Premature age-adjusted mortality rate per 100,000 population
Covariate: \(Urbanicity_n\)
- Six categories (same as Aim 1): large central metro, large fringe metro, medium metro, small metro, micropolitan, noncore
Spatial error: \(u_{it}\)
- Accounts for spatial correlation using Queen’s contiguity
Aim 3: Results
Connection to county-level health
\[ y_{it} = \beta_0 + \sum_{r=1}^{3} \beta_{o_r} o_{ir} + \sum_{r=1}^{3} \beta_{d_r} d_{jr} + \sum_{r=1}^{3} \beta_{t_r} t_{kr} + \sum_{n=2}^{6} \gamma_n \text{Urbanicity}_n + \mu_{it} \]
Increased origin loadings in Component 2: Specific phenomena \((o_{i2})\) are associated with higher premature mortality rates
Increased destination loadings in Component 1: Urban-to-rural \((d_{i1})\) and Component 3: State borders \((d_{i3})\) are associated with lower premature mortality rates
Even after accounting for a county’s position within a migration system, the effect of urbanicity remains significant
- Urbanicity gradient: effect of urbanicity increases as counties become more rural (i.e. as \(n\) increases)
Aim 3: Pitfalls and Alternatives
Somewhat dependent upon individual perception
Non-convex
- Ensemble decomposition and/or consensus clustering
IRS data
Highly sparse
Temporality issues
Non-representative
NARROWED GEOGRAPHICAL SCOPE
Conclusions
Aim 1:
Contextual factors persist even after we adjust for compositional factors.
Contextual factors matter more in urban counties than in rural counties.
Aim 2:
There is some evidence of health-related selection occurring for migration from urban origins to urban destinations.
No evidence of health-related selection occurring for migration to rural destinations.
Aim 3:
The effect of urbanicity persists even after we adjust for a county’s position within a migration system.
Tensor decomposition is an exciting new method for distilling the complexities of migration.
Future Questions
Can we replicate the findings of Aims 2 and 3 using the ACS five year data?
Need for smaller unit of analysis (census tract? mobile device?) and more inclusive data OR state and regions specific analyses
What will happen to places that experience forced or voluntary exodus of unauthorized immigrants?
Increased need for place-based assessments of health to quantify potential impacts of the next administration
Technical acknowledgements
I acknowledge the use of ChatGPT for code generation, editorial writing, and error correction
Many thanks to the UW-Madison Writing Center
This research would not have been possible without resources available through the Center for High Throughput Computing
This work is dependent upon amazing open source tools and templates
Personal acknowledgements
CHRR folks
Mom, Dad, Erik, Andrew
My committee members: Amy, Marjory, Shaneda, Paul, and Jenna
Everyone here today - THANK YOU!
Extras
Aim 1: Results
Aim 1: Results
Aim 1: Results
Aim 2: Data
All analyses are from the perspective of a destination county of migration. Counties are included in our analyses if they:
Are a migration destination
Are part of the contiguous US
Included a total of 3107 US counties in our analyses.
Used years 2011 to 2019 (9 years total)
Aim 2: Variables - derivation of \(d_{ij}\)
Exploratory explanatory factor: adjusted migration term, denoted \(amig_{it}\) for county \(i\) in year \(t\) with new parameters \(k_{ij}\) and \(l_i\) for each origin-destination pair.
\[ amig_{it}(k_{ij}, l_i) = \frac{ \sum_{j\ne i} out_{jit} ( y_{j,t-1} + k_{ij}) + (y_{i,t-1} + l_i) (pop_{i, t-1} - \sum_{j\ne i} out_{ijt})}{ \sum_{j\ne i} out_{jit} + (pop_{i,t-1} - \sum_{j\ne i} out_{ijt})} \]
When \(k_{ij} <0\) : movers from county \(j\) to county \(i\) are healthier, on average, than the typical person in their origin county \(j\)
When \(k_{ij} >0\) : movers from county \(j\) to county \(i\) are less healthy, on average, than the typical person in their origin county \(j\)
When \(l_i < 0\) : stayers in county \(i\) are healthier, on average, than the typical person in county \(i\).
When \(l_i > 0\) : stayers in county \(i\) are less healthy, on average, than the typical person in county \(i\).
Aim 2: Variables
Exploratory explanatory factor: adjusted migration term, denoted \(amig_{it}\) for county \(i\) in year \(t\)
Let \(d_{ij} = k_{ij} - l_i\). Then:
\[ amig_{it} = \frac{ \sum_{j\ne i} out_{jit} ( y_{j,t-1} + l_i + d_{ij}) + (y_{i,t-1} +l_i)( pop_{i, t-1} - \sum_{j\ne i} out_{ijt})}{ \sum_{j\ne i} out_{jit} + (pop_{i,t-1} - \sum_{j\ne i} out_{ijt})} \]
Aim 2: Variables
Exploratory explanatory factor: adjusted migration term, denoted \(amig_{it}\) for county \(i\) in year \(t\)
\[ amig_{it} = \frac{ \sum_{j\ne i} out_{jit} ( y_{j,t-1} + d_{ij}) + y_{i,t-1} (pop_{i, t-1} - \sum_{j\ne i} out_{ijt})}{ \sum_{j\ne i} out_{jit} + (pop_{i,t-1} - \sum_{j\ne i} out_{ijt})} +l_i\]
Aim 2: Results
Descriptive Statistics
Aim 3: Results
Connection to county-level health
